aontu 0.52.1 → 0.54.0

package/src/provenance.ts

@@ -0,0 +1,437 @@
+/* Copyright (c) 2025 Richard Rodger, MIT License */
+
+// THE PROVENANCE RECORDER (G7 phase 3,
+// docs/capability-review/g7-machine-access.md): what CONTRIBUTED to
+// the value at a path, in order, with the site each contribution was
+// written at. `why` is the positive twin of G2's error report — errors
+// explain what failed to unify, this explains what did.
+//
+// The recorder rides the CONTEXT and is off by default: `unite` pays
+// one property load on the normal path, and an instrumented run pays
+// site materialisation knowingly. It records at `unite` and nowhere
+// else, because that is the one place every meet passes through — the
+// same reason G3's deprecation rider lives there. A meet is where
+// information currently vanishes, so a meet is what to record.
+//
+// Contributions are the operands that were NOT produced by an earlier
+// meet at the same path: a value that a previous meet made is an
+// intermediate, not a source. Deduplicated by (path, val id) — ids are
+// unique per run — so fixpoint revisits across up to `maxcc` passes do
+// not multiply the record.
+//
+// Roles come from the operand itself, with no further instrumentation:
+// a reference is still a RefVal when it meets its peer (its canon is
+// the path it names), a preference is still a PrefVal (its canon is
+// the `*` form), and a spread-applied template is marked once where
+// the spread is applied. Precedence is spread > ref > pref > literal:
+// a preference INSIDE a spread template is a spread contribution,
+// which is what the author needs to be told.
+
+
+
+export type WhyRole = 'literal' | 'spread' | 'ref' | 'pref'
+
+// The G2 site object, minus its data/schema role: a contribution's
+// role is its own (above), and a `why` run has one document.
+export type WhySite = {
+  col: number
+  file: string
+  // The extent in UTF-16 code units, or -1 when unknown. The same
+  // field, and the same meaning, as VetSite.len (ts/src/vet.ts).
+  len: number
+  row: number
+}
+
+export type WhyConjunct = {
+  canon: string
+  role: WhyRole
+  site: WhySite
+  // The SOURCE TEXT this contribution was written as.
+  //
+  // `canon` is the value; `src` is the spelling. They are not the same
+  // thing, and the difference is the whole reason this record exists:
+  // `port: 0x1F` contributes canon `31` from source `0x1F`, so a reader
+  // told only the canon cannot find, verify or replace what was
+  // actually written. Empty when the contribution occupies no source —
+  // a value unification minted rather than a document wrote.
+  src: string
+  // The PREFERENCE RANK, 0-based, when the contribution is one: `*x`
+  // is 0, `**x` is 1. Absent for anything else. The engine's own
+  // number (PrefVal.rank), so a reader arbitrating between ranked
+  // contributions -- the meet ladder -- need not count stars in a
+  // canon string.
+  rank?: number
+}
+
+export type WhyRecord = {
+  conjuncts: WhyConjunct[]
+  path: string
+  value: string
+}
+
+
+// Set on a spread template's per-key clone, at the one place a spread
+// is applied (MapVal/ListVal.unify). Nothing else reads it.
+export const FROM_SPREAD = '_fromSpread'
+
+
+// THE AUTHORED MARK, and it is a property of the VALUE rather than a
+// set of ids held beside it (the review's finding E). A contribution
+// is a value the author wrote, and the recorder used to decide that by
+// looking the operand's id up in a set stamped over the parsed tree --
+// which is true of the parsed tree and of nothing derived from it. So
+// every value that reached a path through a CLONE was dark: a default
+// flowing into a `pack()`-generated child, a shape carried by a `$ref`,
+// a value a spread template stamped. `why` answered "(no contributions:
+// nothing met at this path)" over a value it had just printed, with
+// exit 0 -- a false statement, and the one an audit surface may not
+// make (use-cases/BUGS.md §23, §24).
+//
+// A clone of a written value IS that written value, re-instantiated
+// somewhere else: it carries the author's site, so it can be pointed
+// at. Val.clone therefore carries this mark, exactly as it carries the
+// site -- provenance is part of the clone contract, not a recorder
+// bolted beside it. Values the engine MINTS (a kind lifted while a
+// disjunct trials its members, a fold's intermediate) are constructed
+// rather than cloned and stay unmarked, which is what keeps the record
+// to what the author can edit.
+//
+// The mark is only ever SET by an instrumented run (`why` calls
+// writtenFrom; nothing else does), so an ordinary evaluation pays one
+// undefined property read per clone.
+export const WRITTEN = '_written'
+
+
+// THE CONTAINER A VALUE IS PART OF, when the two stand at the SAME
+// path: a junction's members, a preference's inner value, a function's
+// arguments, an operator's operands. `*1|integer` is one thing the
+// author wrote, and `*1` is not a second thing beside it.
+//
+// The recorder already knew that where the container itself met
+// something -- contributing it marked its members `inside` and the
+// filter dropped them. But the container only appears as an operand
+// when the fixpoint happens to meet it whole; where it resolves
+// member-wise instead (a later sibling of a spread, a default reaching
+// a generated child) the members arrived alone and the record split
+// one written value into two contributions at two columns. Which
+// happened was decided by evaluation order, so `why` answered
+// differently for identical statements (use-cases/BUGS.md §22).
+//
+// So the relation is recorded as a fact about the DOCUMENT, at
+// stamping time, and an operand is reported as the outermost written
+// value it is part of. NOT set for a bag's children (they stand at
+// their own, deeper paths) nor for a conjunct's terms (a conjunct is
+// the statement that several things must all hold, and each term is
+// one of them -- which is exactly what the author needs shown).
+export const INNER_OF = '_innerOf'
+
+
+// Mark a spread clone and everything inside it, so a contribution
+// several levels down a template is still known to have come from the
+// template. ONLY on an instrumented run: the walk is O(template) per
+// key per pass, which is real money on a large model and buys nothing
+// when no one is recording.
+//
+// THE GUARD IS A CYCLE GUARD, NOT A "DONE" FLAG, and that distinction
+// is the whole of the review's finding E for sibling position. A
+// template is applied once per destination, and the fixpoint advances
+// values IN PLACE between those applications (AGENTS.md, the mutation
+// caveat): by the time the second key is spread, the template's
+// `replicas` child is no longer the disjunction the first key saw but
+// the value that meet produced. Skipping the walk because the
+// CONTAINER was already marked left every one of those replacements
+// unmarked, so `why` at the first sibling reported the written
+// `*1|integer` as one contribution and at the second reported `*1` and
+// `integer` as two -- identical statements, different answers, decided
+// by which key the fixpoint reached first (use-cases/BUGS.md §22).
+// Marking is idempotent, so re-walking costs a pass and changes
+// nothing where nothing moved.
+export function markSpread(v: any, seen?: Set<any>): void {
+  const marked = seen ?? new Set<any>()
+  if (null == v || true !== v.isVal || marked.has(v)) {
+    return
+  }
+  marked.add(v)
+  v[FROM_SPREAD] = true
+  if (true === v.isMap && null != v.peg) {
+    for (const k of Object.keys(v.peg)) {
+      markSpread(v.peg[k], marked)
+    }
+  }
+  else if (true === v.isList && null != v.peg) {
+    for (const k of Object.keys(v.peg)) {
+      markSpread(v.peg[k], marked)
+    }
+  }
+  else if (Array.isArray(v.peg)) {
+    // A junction, a func's arguments, an op's operands: every one of
+    // them can hold the value that reaches the destination.
+    for (const m of v.peg) {
+      markSpread(m, marked)
+    }
+  }
+  else if (true === v.isPref) {
+    markSpread(v.peg, marked)
+  }
+}
+
+
+// The children that stand at the SAME path as v. See INNER_OF.
+function samePathKids(v: any): any[] {
+  if (true === v.isMap || true === v.isList || true === v.isConjunct) {
+    return []
+  }
+  if (true === v.isPref) {
+    return null != v.peg && true === v.peg.isVal ? [v.peg] : []
+  }
+  return Array.isArray(v.peg)
+    ? v.peg.filter((k: any) => null != k && true === k.isVal) : []
+}
+
+
+// Order contributions the way the document reads: by file, then row,
+// then column, with the canon as the last tiebreak so the order is
+// total even for two values written at the same position (which a
+// merged duplicate key can produce).
+function cmpSite(a: Contribution, b: Contribution): number {
+  return a.site.file.localeCompare(b.site.file) ||
+    a.site.row - b.site.row ||
+    a.site.col - b.site.col ||
+    a.canon.localeCompare(b.canon)
+}
+
+
+function roleOf(v: any): WhyRole {
+  if (true === v[FROM_SPREAD]) {
+    return 'spread'
+  }
+  if (true === v.isRef) {
+    return 'ref'
+  }
+  if (true === v.isPref) {
+    return 'pref'
+  }
+  return 'literal'
+}
+
+
+type Contribution = WhyConjunct & {
+  id: number
+}
+
+type PathRecord = {
+  conjuncts: Contribution[]
+  // Ids of values PRODUCED by a meet at this path: an operand among
+  // them is an intermediate result, not a source contribution.
+  made: Set<number>
+  seen: Set<number>
+}
+
+
+export class Provenance {
+  paths: Map<string, PathRecord> = new Map()
+
+  // id -> the written container it names, for INNER_OF. The mark on a
+  // value is the container's ID rather than the container itself: a
+  // Val holding another Val as an own property is a reference cycle
+  // through the tree, and enough of the engine walks a value's own
+  // properties that one is not safe to introduce.
+  containers: Map<number, any> = new Map()
+
+  // Stamp the parsed tree with the AUTHORED mark: everything the
+  // author wrote, before unification starts. A value minted during
+  // unification — a kind lifted from a leaf while a disjunct trials
+  // its members, a fold's intermediate — is the engine's own work, not
+  // a contribution the author can be pointed at. A CLONE of a marked
+  // value keeps the mark (see WRITTEN above), because it is the same
+  // written value somewhere else. Called once, before unify, by `why`.
+  writtenFrom(v: any): void {
+    if (null == v || true !== v.isVal || true === v[WRITTEN]) {
+      return
+    }
+    v[WRITTEN] = true
+    const kids: any[] =
+      (true === v.isMap || true === v.isList) && null != v.peg
+        ? Object.keys(v.peg).map((k) => v.peg[k])
+        : Array.isArray(v.peg) ? v.peg
+          : null != v.peg && true === v.peg.isVal ? [v.peg]
+            : []
+    for (const k of kids) {
+      this.writtenFrom(k)
+    }
+    // OUTERMOST WINS: the walk is top-down, so a value already pointed
+    // at a container is inside that one and this one, and the answer
+    // the author wants is the whole written statement.
+    for (const k of samePathKids(v)) {
+      if (null == k[INNER_OF]) {
+        k[INNER_OF] = v.id
+        this.containers.set(v.id, v)
+      }
+    }
+    if (null != v.spread?.cj) {
+      this.writtenFrom(v.spread.cj)
+    }
+  }
+
+  // One meet. Both operands are candidate contributions; the result is
+  // remembered so a later meet does not mistake it for a source.
+  record(path: string[], a: any, b: any, out: any): void {
+    const key = path.join('.')
+    let rec = this.paths.get(key)
+    if (null == rec) {
+      rec = {
+        conjuncts: [], made: new Set(), seen: new Set(),
+      }
+      this.paths.set(key, rec)
+    }
+
+    this.contribute(rec, a)
+    this.contribute(rec, b)
+
+    if (null != out && true === out.isVal && out !== a && out !== b) {
+      rec.made.add(out.id)
+    }
+  }
+
+  private contribute(rec: PathRecord, v: any): void {
+    // TOP is the unit element and a nil is a failure, neither of which
+    // is information the author wrote. A value an earlier meet MADE is
+    // an intermediate; the source that made it is already recorded.
+    if (null == v || true !== v.isVal || true === v.isTop || true === v.isNil ||
+      rec.made.has(v.id) || rec.seen.has(v.id)) {
+      return
+    }
+
+    // PART OF a written value is not a value beside it: report the
+    // whole statement the author wrote, whichever piece of it the
+    // fixpoint happened to meet here. See INNER_OF.
+    let outer: any = v
+    for (let up = this.containers.get(outer[INNER_OF]);
+      null != up; up = this.containers.get(outer[INNER_OF])) {
+      outer = up
+    }
+    if (outer !== v) {
+      this.contribute(rec, outer)
+      return
+    }
+    // Not the author's: see WRITTEN.
+    if (true !== v[WRITTEN] && true !== v[FROM_SPREAD]) {
+      return
+    }
+    // A CONJUNCT is not one contribution, it is the statement that
+    // several must all hold — duplicate keys merged at parse, an
+    // explicit `a & b`. Its own site is nowhere (the merge has no
+    // source position), while its terms each have one, which is what
+    // the author needs to be shown.
+    if (true === v.isConjunct && Array.isArray(v.peg)) {
+      rec.seen.add(v.id)
+      for (const term of v.peg) {
+        this.contribute(rec, term)
+      }
+      return
+    }
+    rec.seen.add(v.id)
+    rec.conjuncts.push({
+      canon: v.canon,
+      id: v.id,
+      role: roleOf(v),
+      // COALESCED, unlike vet's siteOf: a `why` run reads whatever
+      // source it was handed, and an inline document (a spec row, a
+      // piped stdin) has no file name to stamp. The Go port answers
+      // the empty string for the same value, so the two agree.
+      site: {
+        col: v.site.col, file: v.site.url ?? '', len: v.site.len,
+        row: v.site.row,
+      },
+      src: v.site.src,
+      ...(true === v.isPref ? { rank: v.rank } : {}),
+    })
+  }
+
+  // THE VALUE THAT STANDS at a path is a contribution when nothing met
+  // there and the author wrote it. A meet is where information
+  // vanishes, so a meet is what the recorder watches -- but a
+  // generator PLACES a value without meeting anything, and `why` then
+  // answered "(no contributions: nothing met at this path)" over a
+  // value it had just printed. That is literally true and practically
+  // false: the author is asking where the value came from, and it came
+  // from somewhere they can be shown (use-cases/BUGS.md §23).
+  //
+  // Only when the record is otherwise EMPTY. Where something did meet,
+  // the standing value is that meet's result -- an intermediate, and
+  // the recorder's oldest rule is that a result is not a source.
+  stands(path: string[], v: any): void {
+    const key = path.join('.')
+    const rec = this.paths.get(key)
+    if (null != rec && 0 < rec.conjuncts.length) {
+      return
+    }
+    this.record(path, v, undefined, undefined)
+  }
+
+  // The record at one path. Empty when nothing met there and nothing
+  // the author wrote stands there either — which is a true and useful
+  // answer rather than an error.
+  //
+  // ONLY WHOLE WRITTEN VALUES are contributions. A Val's own unify
+  // re-enters `unite` at the same path — a disjunct trials each member
+  // there, a constraint meets its atoms there — and those members are
+  // PARTS OF one written value, not further values beside it. That is
+  // settled BEFORE a member is ever pushed, by the INNER_OF fact
+  // stamped over the document (see `contribute`), which is why no
+  // filter runs here: a per-path "inside" set used to do it, and it
+  // could only work where the container itself happened to meet
+  // something at the same path — the order-dependence finding E
+  // records.
+  at(path: string[]): WhyConjunct[] {
+    const rec = this.paths.get(path.join('.'))
+    if (null == rec) {
+      return []
+    }
+    // SOURCE ORDER, not meet order: the two are the same in simple
+    // cases and diverge with the fixpoint's fold order, which is an
+    // engine detail and a parity risk. Sites are parse data, identical
+    // in both ports, so ordering by them makes the record read as the
+    // document reads and pins it across implementations.
+    // ONE WRITTEN TOKEN IS ONE CONTRIBUTION, and the SITE is what
+    // identifies it -- not the val id, and not the canon.
+    //
+    // Not the id, because provenance travels through clones now: a
+    // written value and a clone of it are the same statement in the
+    // same place, and a path that met both would list it twice.
+    //
+    // Not the canon, because the same written value reaches a path at
+    // different stages of narrowing -- `3|(1|2)` as the author wrote
+    // it and `3|1|2` after a fold -- and both name one token.
+    //
+    // Not the role either: the role says how the value REACHED this
+    // path, not which value it is, and one written value can reach a
+    // path both ways (a template applied to a key whose value is also
+    // written there). Keeping the literal would throw away the more
+    // informative half, so the roles have a precedence.
+    //
+    // ONLY WHERE THE SITE IS REAL. An unsited contribution (row -1)
+    // cannot be told apart from another unsited one, so those are kept
+    // as they come rather than collapsed into whichever arrived first.
+    const shown = new Map<string, Contribution>()
+    const order = ['spread', 'ref', 'pref', 'literal']
+    const out: Contribution[] = []
+    for (const c of rec.conjuncts.slice().sort(cmpSite)) {
+      if (c.site.row < 0) {
+        out.push(c)
+        continue
+      }
+      const key = [c.src, c.site.file,
+        c.site.row, c.site.col, c.site.len].join('\u0000')
+      const had = shown.get(key)
+      if (null == had) {
+        shown.set(key, c)
+        out.push(c)
+      }
+      else if (order.indexOf(c.role) < order.indexOf(had.role)) {
+        had.role = c.role
+      }
+    }
+    return out.map(({ id, ...rest }) => rest)
+  }
+}